Deep Visual Saliency on Stereoscopic Images.

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Visual saliency in stereoscopic 3D (S3D) images is significantly affected by image quality.
• Existing saliency detection algorithms often fail with distorted images, hindering applications like image quality prediction and restoration.
• Predicting the nonlinear relationship between image quality and visual saliency remains a challenge.

Purpose of the Study:

• To develop a deep learning model for accurate and reliable visual saliency prediction in S3D images, even with distortions.
• To investigate the role of low-level features in adapting saliency detection to human perception.
• To improve image quality prediction, restoration, and discomfort reduction in S3D content.

Main Methods:

• Proposed a deep learning scheme named Deep Visual Saliency (DeepVS).
• Utilized seven low-level features (contrast, luminance, depth) derived from S3D image pairs.
• Employed regression and fully convolutional neural networks (CNNs) to model human visual saliency.

Main Results:

• Low-level features were found to effectively extract distortion and saliency information.
• The DeepVS model demonstrated accurate saliency prediction capabilities in the presence of image distortions.
• Predicted saliency maps showed up to 70% correlation with human gaze patterns.

Conclusions:

• Hand-crafted features are crucial inputs for deep neural networks in S3D saliency detection.
• The DeepVS model offers a more reliable approach to predicting visual attention in S3D images.
• This research contributes to advancements in S3D image processing and understanding human visual perception.